Q&A: Ian Beaver, Chief Scientist at Verint - Next IT, on Conversational Analysis
Conversational analysis can provide a lot of valuable information for companies, as long as they know how to use it. James Larson, co-chair of the SpeechTEK conference talked to Ian Beaver, chief scientist at Verint - Next IT, about the technology, which will be part of his SpeechTEK 2020 presentation.
What is conversational analysis?
Conversational analysis is part of the formal field of discourse analysis. In the business application, conversational analysis is the act of reviewing text or speech logs containing conversations between humans and machines or humans and other humans for the purpose of understanding the topics of interest, intentions, context, and motivations of customers or constituents. This is accomplished with the help of various natural language processing (NLP) tools that have been developed specifically for this purpose. These tools allow analysists to quickly surface these features in the large volumes of data that we see today.
Analysts review logs from a specific source or language domain, such as finance or insurance, to look for patterns within the conversations that can indicate actions to be taken by the business or government agency. These logs can originate from any interface, such as email, forums, chat, virtual assistants, call recordings, social media, etc., that allows for natural language communication with customers, For example, if a competitor releases a new product, an analyst might discover that some customers are increasingly asking customer service reps how their company's product compares to the competitor's. This insight on the lack of customer awareness might lead to actions such as publishing a white paper comparing the products, educating contact center agents on the benefits of the company product versus the competitors, or adding a comparison to the company web page. These kind of insights allow companies to react to behavioral and motivational changes in their customer bases, increasing quality of self-service and customer service, which in turn increases customer satisfaction.
What are the major challenges in conversational analysis?
One of the challenges in analysis of conversation is the tools need to be built for or adapted to each domain or language you are attempting to analyze. For example, if a tool used to identify topics was created from public data sources such as Wikipedia, it might not be able to identify topics in conversations specific to a company's products or services. Much of the company-specific terminology (e.g. product names or service codes) will either not appear in Wikipedia, or worse, will have different meanings. Therefore, we spend a lot of time developing analysis tools that can be quickly adapted to specific organizations or language domains.
A second challenge is the volume of data that is generated across all of the channels a specific company might deploy. We have customers that see millions of conversations per month over just a single channel. Unlike analytics that is more focused in aggregating and trending events, we are looking much deeper into the motivations of the participants. We need to answer questions like: what was the user wanting to accomplish verses what they were able to accomplish? What was the real intention by this utterance or input within the context of this conversation? Were there any misunderstandings in this conversation and why did they occur? These types of questions are difficult to answer at scale without specialized tools leveraging many areas of artificial intelligence, such as NLP and machine learning, and research in behavioral psychology and communication.
How can analyzing a conversation log help improve customer service?
Improving customer service is the iterative process of understanding which tasks customers are trying to accomplish, then for each task determining the most efficient means to complete it, and finally creating or improving those means for the customers. Customer service extends even to internal employee interactions, such as handling IT help desk tickets or responding to open enrollment questions in HR. Efficiency can be measured by the time or effort it takes for a customer to reach resolution, or by how much money or effort is spent by the company to resolve the issue.
Human or automated customer service agents on calls, over chat, or through other channels must first determine the customer intention, and can then perform actions on their behalf to help them accomplish it. Conversational analysis comes into play to find shortcomings or inefficiencies in these customer interactions that can then be used to help companies and agencies quantify issues and prioritize where improvement should be made. The outcome of conversational analysis is a prioritized list of tasks that are ripe for automation, areas where there are knowledge gaps, such as missing content on websites, and workflow improvements for existing tasks. The analysis can only provide recommendations; companies then need to act on them to complete the customer service improvement cycle.
What are the limits to which a company can use data extracted from a conversation log? Should a company be allowed to search a conversation log to uncover items that it might sell to its users? Can a company extract information from a conversation log to determine a customer's credit rating?
Data usage rights and how it impacts user privacy is a topic that is under much deliberation today. With recent legislation such as Europe's General Data Protection Regulation and the California Consumer Privacy Act, companies have to re-evaluate what they can do with user-generated data while still being able to understand what their customers want from them. In general, companies can use customer service data to improve their own products and services provided the customer is notified that the data may be used in that manner and the company provides means to let customers opt-out or manage any personally identifying data.
Even with such constraints, it is still possible to perform conversational analysis. Many privacy-preserving methods have been developed, such as differential privacy, which injects random noise into the data making it nearly impossible to identify the originating person. Personally identifiable data can also be scrubbed prior to performing analysis. These methods, when applied properly, allow a company to perform data analysis and remain compliant. These privacy laws are concerned with the identification of an individual or household, while conversational analysis is about understanding the topics of interest and intentions of customers that can be discovered without identifying them. If someone calls to complain about not finding information needed on the company website, we can determine the company website has a knowledge gap without knowing who called.
In light of this, mining conversations to discover items that could be sold to a customer base is permissible as long as measures are taken to protect customer privacy. For example, you could determine that 30 percent of your customer base would prefer that you offered product X from analyzing your customer feedback data as long the data was anonymized. However, you could not determine a customer's credit rating from conversations without prior consent as that would require extracting personally identifying information such as age or location.
Is there a shortage of professionals who understand data analysis?
In my experience, there has been a growing pool of young professionals entering the workforce with educations focused on data science. More and more universities are creating data science programs and graduate degree options to meet the demands of industry. Where I have seen shortages is in experienced analysts who have developed intuition from years of experience in the trenches as opposed to just having familiarity with the tools and topics from a formal education There is the saying that if you torture the data long enough it will confess to anything. Having the intuition to know what the data is really saying and where some latent bias might be affecting your results is developed over time as practitioners are faced with real-world problems.
Can analytics be applied to multiple data sources, such as call center data and social media data?
Absolutely. Combining these data sources gives a wholistic view of customer service, as single channels can be biased towards the demographics that use that channel and issues with one channel can be presented on another (I tried searching for this on your website but…). In information security there is the concept of the attack surface, which is the sum of points an attacker can enter or extract data from an organization. A basic security measure is then to audit all of the points that make up this surface and ensure they are secure. With customer service, organizations need to audit all of the points that make up their customer-facing surface to ensure they are performing optimally and that the organization is meeting the needs of customers regardless of which channel they use.
How can companies deal with data that contains errors?
Conversational data is an immutable record of who said what to whom and when. Therefore, once it is recorded it does not change, so the only real errors present would be either created in translation (speech-to-text) or in communication between participants. Conversational analysts often look to uncover such communications errors by looking for conversational cues that humans pass to each other to signal that the conversation is not progressing as it should For instance, if a user begins an utterance with correctional language, such as"No, I said&hellip" or repeats or rephrases a previous utterance, we know there is a communication error occurring and we should inspect the speech translations, if present, and agent responses to determine the cause.
In the case of automated or self-service, errors or inefficiencies in workflows can surface through the types of clarifying questions people ask (When the form says enter X, does it mean … ?) We can discover these questions and use them to provide recommendations to correct these workflows, eliminating future clarifying customer contacts and improving customer experience.
To see presentations by Ian Beaver and other speech technology experts, register to attend SpeechTEK 2020 today.